[time-nuts] Divide by five

[Magnus Danielson]

Poul-Henning,

On 11/09/2014 09:44 PM, Poul-Henning Kamp wrote:
> --------
> In message <CABbxVHtommjwSWq1i=oH-u1S=G6P=xu8E0YzeKaDg-VchgkGwQ at mail.gmail.com>
> , Chris Albertson writes:
>
>> NTP does not pick the best clock.   NTP finds
>> the subset of clocks that track each other.
>
> NTP does indeed find the best clock from the subset of clocks
> which pass its sanity check, and then it uses only that one.
>
> There are several problems with that, and as we speak I'm developing
> a new algorithm which at least so far, gives much superior performance.
>
> You can read my musings about this here:
>
> 	http://phk.freebsd.dk/time/20141107.html
>
> My goal is to release the new NTP client before X-mas.
>

Just as you point out, there is no easy way around the fact that delays 
may be asymmetric. When one analyzes the problem, the actual time-error 
and delays in both directions cannot be solved with the two measurements 
one do, three unknowns and two equations. The sum of the observations, 
forming the Round-Trip-Time, is however the only value we can trust, and 
worst case asymmetry will be that of the RTT, regardless of statistical 
distribution. Using RTT as a worst-case estimator is thus enough if you 
don't know better, and using RTT as a simple stability estimator isn't 
all that bad for a simple system.

The NTP scatter-plot wedge is another way to present it, and finding the 
"tip" of that wedge is really about finding the min-value of delay in 
each direction prior to doing the two-way time-transfer equations.

Cheers,
Magnus